Money and Politics: The Effects of Campaign Spending Limits on Political Entry and Competition. Carlos Varjão UC Berkeley. December 2017.

Transcription

1 Money and Politics: The Effects of Campaign Spending Limits on Political Entry and Competition Eric Avis Claudio Ferraz Frederico Finan Carlos Varjão UC Berkeley PUC-Rio UC Berkeley Stanford December 2017 Abstract This paper examines the effects of campaign spending limits on political entry and competition. We study a reform in Brazil that imposed limits on campaign spending for mayoral elections. These limits were implemented with a discontinuous kink which we exploit for causal identification. We find that stricter limits increase political competition by creating a larger pool of candidates that is on average less wealthy. Stricter spending limits also reduce the likelihood that mayors are reelected. We interpret our reduced-form findings using a contest model with endogenous entry of candidates. Keywords: Political Entry, Political Selection, Campaign Finance, Campaign Contributions, Incumbency Advantage Department of Economics, Evans Hall, Berkeley, California eavis@econ.berkeley.edu Department of Economics, Pontifícia Universidade Católica do Rio de Janeiro (PUC-Rio), Rua Marquês de São Vicente, 225- Gávea Rio de Janeiro, RJ, , Brasil. cferraz@econ.puc-rio.br; and BREAD Department of Economics, Evans Hall, Berkeley, California ffinan@berkeley.edu; and BREAD, IZA, NBER Graduate School of Business, 655 Knight Way, Stanford, California cvarjao@stanford.edu

2 1 Introduction Among the many factors critical for a properly functioning democracy, few have been as widely debated as campaign financing. For some, money in politics serves as an expression of free speech and an effective instrument for informing voters and building an inclusive democracy. For others, the unrestrained use of money in politics can erode the functioning of democracy as it can lead to excessive campaigning, unequal access to power, and politicians who are beholden to special interest groups. 1 In practice, almost every country with political pluralism has adopted some type of political finance regulation ranging from information and disclosure requirements to limits on campaign contributions and/or expenditures (Scarrow, 2007). Countries such as Canada and the UK have been limiting campaign spending by parties and individuals for many decades. 2 More recently Belgium, Chile, France, Israel, New Zealand, South Korea and many others have also adopted campaign spending caps in order to limit the role of money in elections. 3 Despite the widespread adoption of spending limits, our understanding of how they impact the political process is limited. As we show theoretically, spending limits can affect both who enters politics and who gets elected. However, because the decision to run for office depends not only on a candidate s own characteristics, but those of his opponents, the effects of spending limits can be ambiguous. Empirically, to estimate the effects of spending limits on political behavior presents some difficult challenges. Campaign finance reform is usually applied uniformly across elections and jurisdictions, which makes it difficult to identify an appropriate comparison group. In addition, few countries provide information on the characteristics and campaign spending of both their elected and non-elected candidates. It is important to have data on both types of candidates if, as theory suggests, spending caps affect not only the identity of who is elected, but also who chooses to run. In this paper, we provide (to our knowledge) the first causal estimates of the effects of campaign spending limits on political entry and selection. We do so in the context of a recent campaign 1 For example, see Coate (2004), Prat (2002), Prat (2006), and Scarrow (2007). 2 Currently, political parties in Canada can spend only 73.5 cents for every voter in districts in which they are competing. In the United Kingdom, legislation regulating expenditures has been in place since the Corrupt and Illegal Practices Prevention Act In the 2005 general election, campaign expenditure at the national level were limited to approximately US$42,000 per constituency contested. 3 Two thirds of the OECD countries have introduced campaign spending limits for parties or candidates (Speck, 2013). One of the few exceptions among rich countries is the U.S. where the Supreme Court ruled mandatory spending limits as an unconstitutional curtailment of free speech. 1

3 finance reform in Brazil. Amid a massive corruption scandal that included the diversion of public funds to political campaigns, the Brazilian Supreme Court in 2015 decided to ban corporate donations and Congress passed a law that imposed campaign spending limits in future elections. The spending caps, which vary by municipality, create a discontinuous kink in the amount candidates can spend in local elections. We exploit this discontinuity together with a rich dataset on all candidates elected and non-elected to explore how spending limits affect the entry decisions of candidates, their characteristics, and electoral results for mayors. Our analysis, which focuses on municipalities near the point of discontinuity, shows that places subject to lower spending limits are more politically competitive (attract more candidates). Our estimates suggest that a 25 percent decrease in spending caps leads to a 9 percent increase in the number of individuals who run for office, and an average candidate who is 40 percent less wealthy. Our results show that spending limits also affect political selection. We find that re-election rates are 11 percentage points lower in places with more stringent spending caps, suggesting that spending limits reduce incumbency advantage. We also find evidence that more stringent spending caps lead to the election of less wealthy candidates, and of candidates who spend less of their own funds in their campaigns. We interpret these reduced-form results using a contest model with endogenous candidate entry. In this model, candidates differ in their effectiveness of spending and their cost of fundraising. Candidates then choose whether to enter the race and how much to spend, taking into account both spending limits and the equilibrium responses of their opponents. Given the best response functions of potential candidates, we then estimate the model via maximum likelihood. The estimation of the model complements our reduced-form analysis in three notable ways. First, we can examine how spending caps affect candidate entry based on a candidate s effectiveness of spending and ability to raise funds: two unobservable traits that are important for computing the elasticity of vote shares with respect to campaign spending and for understanding the welfare effects of spending caps. Second, our reduced-form finding that lower spending caps decrease the likelihood of being reelected can be explained by two factors: a reduction in spending that disproportionately affects incumbents more than challengers, or the entry of new candidates. By estimating the model, we can compute the relative contributions of these two channels, and thus assess from a policy perspective whether spending caps can affect reelection rates, even in environments with high barriers to entry into politics. Finally, whereas the reduced-form analysis is restricted to the estimation of a localized effect at the threshold, the structural estimates allow us to consider the effects of adopting different spending limits. 2

4 Our model fits the data well and reproduces all of our quasi-experimental findings. Our model s estimates suggest that fundraising tends to be less costly for incumbents than for challengers, indicating that if all else were equal, incumbents will raise and spend more money in equilibrium than challengers. Incumbents also tend to be more effective spenders than challengers, implying that if all candidates within a municipality spent the same amount on campaigning, the incumbent would be the most likely to win. Interestingly, we also find a significant positive correlation between effectiveness in spending and cost of fundraising for both incumbents and challengers. Candidates who are likely to spend the most i.e. the candidates with the lowest fundrasing costs will be candidates who, if every candidate spent the same amount, attract fewer votes. From these estimates, we also compute the elasticity of vote share with respect to campaign spending. We find that on average, increasing campaign spending by 1 percent results in an increase of 0.74 percent in the vote share of challengers, compared to 0.64 percent for incumbents. These numbers translate to a marginal cost of a vote of R$22 ($6.6) for the average challenger and R$27 ($8.1) for the average incumbent in our sample. The larger elasticity estimate for challengers is consistent with the extensive literature evaluating the returns to campaign spending (e.g. Gerber (2004)). In contrast to previous studies, a key driver to our result is the unobserved heterogeneity in fundraising abilities, which leads incumbents to invest a higher share of the total expenditures in the election. Hence, the lower incumbent elasticity is the result of equilibrium expenditure decisions, rather than the consequence of a fundamental ineffectiveness in incumbent campaigning. These differences in unobserved heterogeneity between incumbents and challengers also have implications for how one should interpret our reduced-form finding that stricter spending caps reduce the electoral performance of incumbents. Because incumbents have to pay more for a vote and are more likely to be constrained by the cap, a stricter cap reduces the incumbent s vote share disproportionately more than those of the challengers, independent of the number of challengers in the race. At the same time, a stricter spending cap increases entry by new challengers, which can also reduce the incumbent s vote share. With our structural estimates, we can distinguish between these two mechanisms. We find that both channels play a role, and that their relative effect sizes depend on the threshold. For example, in our simulations, lowering the spending cap from R$100,000 to R$50,000 reduces the average incumbent vote share by 6 percentage points. But if entry of new candidates is restricted, as is the case in a 2-party electoral system, then the lowering of the cap would only lead to 3 percentage points decline. The entry effect becomes relatively more important as the spending cap is decreased: reducing the limit to R$25,000 from R$100,000 implies a 14 percentage points 3

5 reduction in incumbent vote share, and 57% of that can be attributed to an entry effect. Overall, our findings highlight that in order to assess the effects of the introduction of campaign spending limits, one must take care to not only consider the equilibrium effects of the policy on the current candidates contesting the election, but also on the entry of new candidates. Our findings contribute directly to a large literature that examines the effects of regulating campaign spending (e.g. Austen-Smith (1987); Prat (2002); Coate (2004); Ashworth (2006)). While there is a rich theoretical literature studying the effects of limits on campaign donations and spending (e.g. Che and Gale (1998); Fang (2002); Pastine and Pastine (2012); Cotton (2012)), empirical studies are rare. The key challenge to causal identification is the potential endogeneity of spending caps. Indeed, while many polities around the world have implemented limits, their presence and magnitude are likely to be correlated with other unobserved factors which also affect potential outcomes. To address this issue, several studies have exploited state-level differences in contribution limits in the U.S. (e.g. Stratmann and Aparicio-Castillo (2006); Barber (2016)). To our knowledge, the only other empirical investigation of the effects of campaign spending limits is Milligan and Rekkas (2008), who study spending caps in Canadian federal elections, and find that higher limits are associated with higher incumbent spending, fewer candidates, and lower voter turnout. Different from these studies, we examine the effects of spending limits on a series of novel outcomes including detailed candidate characteristics for both the elected and non-elected. Furthermore, our research design has the advantage that it requires weaker assumptions to identify the causal effect of spending limits. Our study also relates to a large literature that examines the effects of campaign spending on electoral outcomes (e.g Levitt (1994); Gerber (1998); Erikson and Palfrey (2000); Da Silveira and De Mello (2011)). The vast majority of these empirical studies estimate this relationship in the reduced-form, while trying to control for the unobserved heterogeneity that affects both spending and vote share. We estimate this relationship using our estimated model which has the advantage that it directly accounts for candidates unobserved heterogeneity both in their effectiveness of spending and ability to raise funds. In this respect, our approach is related to Bombardini and Trebbi (2011) who compute the elasticity based on a bargaining model involving politicians and interest groups. Moreover, even though the effect of campaign spending on electoral outcomes is an important input into the design of any campaign finance reform, it is only one piece of the puzzle. As our findings highlight, campaign finance affects not only who enters politics but how many. Accounting for these entry decisions thus becomes critical when using any estimate of the effects of campaign spending on electoral outcomes to guide policy. 4

6 Finally, our work also speaks to research on the identity of politicians and whether limits to campaign spending might level out the playing field between richer and poorer candidates. There is a growing literature following the citizen-candidate models of Osborne and Slivinski (1996) and Besley and Coate (1997) suggesting that identity matters for policy implementation (e.g. Chattopadhyay and Duflo (2004); Besley et al. (2011); Corvalan et al. (2016)). In countries where inequality is high, access to political power might be easier for richer candidates and this might have direct consequences on who gets elected and which types of policies are implemented. Our work suggests that imposing spending caps reduces the average wealth of candidates that run for and are elected as mayor. The rest of the paper is organized as follows. Section 2 describes Brazil s campaign financing laws and presents the data used in the empirical analysis. Section 3 presents the theoretical framework. Section 4 discusses our research design and in Section 5 we present our reduced-form findings, as well as our estimates of the model. Section 6 concludes. 2 Background and Data In this section, we describe campaign financing in Brazil and the 2015 campaign financing law, which was established in response to the Car Wash operation that uncovered one of the world s largest political corruption scandals. The law introduced limits on how much candidates from different municipalities can spend. These differential spending limits form the basis of our identification strategy. We then discuss our data, and present some basic descriptive statistics. 2.1 Municipal Elections and Campaign Financing Local elections in Brazil are held every four years, with the most recent election taking place in October of Candidates need to be registered as a member of a political party in order to run for a political office. The elections are held to elect a municipal mayor and a local council. For municipalities with less than 200,000 registered voters, which represents 98.3 percent of all municipalities, mayors are elected based on simple plurality. For municipalities with 200,000 or more registered voters, candidates for mayor must be elected with at least 50 percent of the votes or a second round runoff is held. Once elected, mayors then face a two-term limit. In contrast, local legislators are elected based on an open-list political representation system, and can be reelected indefinitely. Mayors are important political figures in Brazil. Each year, municipalities receive 5

7 millions of dollars from federal and state governments to provide basic public services such as primary education, health care, and sanitation. The mayor is the agenda setter in how the resources are spent and allocated. 4 Political parties are financed yearly by private contributions and public funds (Fundo Partidário), which are distributed based on the share of votes a party received in the previous election for Congress. All private contributions have to be made prior to the elections to either the political party or directly to an individual candidate. Donations to the parties can be then redistributed to individual candidates. Individuals are allowed to contribute up to 10% of their annual income, unless contributing to their own campaign, in which case there are no limits. Prior to 2015, corporations could contribute up to 2% of gross annual revenues, and there were no restrictions on either total contributions or total campaign spending. Also, Political Action Committees do not exist in Brazil. Campaign spending in Brazil has to be made by individual candidates or political parties on their behalf. Similar to the U.S., both street campaigns and media ads are important forms of campaigning. But different from the U.S., candidates do not need to buy time on TV or radio. In Brazil, TV and radio ads are free and air at predetermined times of the day as determined by Brazil s electoral law. Airtime is distributed according to the share of votes that the candidate for mayor s coalition has in Congress (see Da Silveira and De Mello (2011)). While airtime is free, candidates do have to spend resources on producing the ads. 2.2 The 2015 Campaign Finance Reform On March 14, 2014, Brazil s Federal Police launched an investigation into a local money laundering scheme involving gas stations. This investigation, entitled Lava Jato, has since become one of the largest corruption scandals in the world as investigators uncovered a large corruption scheme involving Petrobras and the largest construction companies in Brazil. Since then, investigators have already uncovered over R$6 billion in paid bribes, charged over 175 people with criminal offenses, and secured 93 convictions. Among those convicted, included key members of Brazil s Workers Party (PT), the PP, and the PMDB who were all found guilty of diverting billions of dollars through procurement contracts to fund their political campaigns. In response to the scandal, Brazil s Supreme Court decided to ban all corporate donations to candidates and parties. This decision led the Brazilian Congress to pass a law on September 2015 that 4 See Ferraz and Finan (2011) for institutional details on Brazil s local politics. 6

8 further limited campaign spending in future elections. 5 The law states that candidates running for mayor are limited to spend the maximum of either R$100,000 (approximately $30,000) or 70% of the highest amount spent by a candidate in the same municipality in the previous election. As stated, the law creates a kink in the amount that candidates can spend at around R$142,858 (70% of R$142,858 is R$100,000.6). For any value lower than R$142,858 the cap is given by R$100,000 while for higher values the cap is given by 70% of the largest value spent in the previous election. The law also stipulated that the caps set by the 70% rule and disclosed in December 2015 should be adjusted by the accumulated inflation between the 2012 and 2016 elections (see Figure 1 for a timeline of the events leading up the 2016 elections). For municipalities capped at R$100,000, they increased the limit by 8.04 percent, which corresponds to the increase in the INPC price index between October 2015 (the month the law was issued) and October For municipalities capped at 70 percent of the maximum amount spent in the 2012 election, the cap was adjusted by 33.7 percent, which corresponds to the increase in the price index that took place between October 2012 and June As a result, the inflation-adjusted caps created a discontinuous kink in the campaign spending limits of about 25 percent, which is what our research design will exploit (see Figure 2). 6 The spending limits apply to any: i) spending made directly by the candidate, ii) spending made by the party on behalf of the candidate, iii) transfers made by the candidate to other candidates (within or across parties) or to political parties, iv) campaign donations estimated in kind or computed as gifts. Candidates that spend more than the limit are subject to severe punishment including a fee of 5 to 10 times the amount that exceeds the limit, ineligibility to run for any political office for 8 years and potential prosecution by electoral courts. Campaign contributions and expenditures are tightly regulated in Brazil. All candidates and parties have to open a bank account exclusively for campaign purposes. All transactions for both contributions and expenditures need to be reported to the Electoral Court within 72 hours and must identify all the entities involved. Every transaction is monitored and made public as soon as the Electoral Court receives the information. 5 See 6 The information on the spending caps is publicly available and can be assessed at the Electoral Court webpage at: 7

9 2.3 Data The data used here come from two sources. The election data come from Brazil s Electoral Commission (TSE). We complement these data with information from the 2010 population census, aggregated at the municipality level. The census data include basic demographic and socio-economic characteristics of the municipality, such as: population size, average income, literary rates, and share of the urban population. The electoral data in Brazil is unusually rich. The data we use in the analysis covers all candidates that ran for mayor in 2012 and In addition to their election results, for each candidate we know a basic set demographic characteristics, such as their gender, age, education level, and self-reported wealth, as well as their campaign contributions and expenditures. Based on this information, we compute at the municipal level, our main political outcomes: campaign spending, the number of candidates that ran for mayor, characteristics of the candidate pool, and re-election rates. Descriptive statistics for the 2016 elections appear in Table 1. On average, elections for mayor attract 3 candidates. Only 13 percent of candidates are female, and only 50 percent of candidates have a college degree. The average candidate in a municipality self-declares asset holdings of about R$1,000,000, but this number masks a lot of heterogeneity as the maximum amount self declared by a candidate in a municipality ranges from R$43,600 to R$24.2 million. In Brazil, incumbents do not enjoy much of an incumbency advantage. Conditional on running for reelection, incumbents were only re-elected in 48.2 percent of municipalities, and received on average 46.8 percent of the votes. In the analysis, we drop open seats (where the mayor is term-limited) so that the sample remains comparable when considering the effects of the spending limits on incumbents. 3 Model Our model builds on the extensive literature studying contests and all-pay auctions in the context of political lobbying and campaigning. 7 In our framework, we extend the n-player contest model with generalized technologies of Cornes and Hartley (2005) in order to incorporate two types of campaign technologies, where one is subject to a cap and the other is not. 8 7 For example, see Tullock (1980), Baron (1994), Siegel (2009) and Jia et al. (2013). For a review of the literature, see Corchón (2007) and Konrad (2009). 8 Although the contest model has not, to our knowledge, been applied to study campaign spending caps, it has been extended to consider the effect of public campaign spending laws (Klumpp et al., 2015). 8

10 We consider an environment in which I 2 candidates compete in an election. Each candidate, indexed by i, chooses how much to spend across two technologies: she chooses an amount x i to spend through formal channels, which is reported to the election commission, and an amount z i to spend through informal channels. Informal spending, which isn t reported to the electoral commission, can include anything from effort spent campaigning on her social media accounts to the use of illicit forms of campaigning, such as vote buying. The candidate s total input into the electoral contest is the weighted sum y i a i x i +b i z i, where a i and b i are measures of each technology s effectiveness in producing votes. We assume that b i < a i for all candidates, so that spending through formal means is more effective. We will refer to a i interchangeably as the campaign effectiveness or popularity of a candidate. After each candidate simultaneously chooses her campaign expenditures, each voter selects his preferred candidate in the election. Voters. We assume there is a continuum of voters who vote sincerely. Each voter s payoff from electing a candidate i is increasing with diminishing returns in the candidate s input into the electoral race. Thus, voters are impressionable" and respond to campaign spending (Grossman and Helpman, 1996). After the candidates have selected their expenditures, an electoral shock ξ in is drawn independently for each voter-candidate pair. Therefore, voter n s utility if he votes for candidate i is v in = log(y i ) + ξ in (1) We normalize the voter s utility to v 0n = 0 if he chooses to abstain. We assume that ξ in are drawn independently from a type I extreme-value distribution, and thus it follows that the share of voters who select candidate i is y i p i = 1 + I k=1 y (2) k A candidate s vote share is given by the share of non-abstaining voters who select that candidate, which is s i = y i I k=1 y. (3) k Politicians. For parsimony, we will assume that candidates seek solely to maximize their expected vote shares net of the costs of campaigning. 9 Normalizing the benefits from the vote share 9 Equivalently, we can assume that politicians seek to maximize the probability of being elected net of the costs of campaigning by letting equation (3) denote the politician s probability of winning the election. 9

11 to 1, we write the candidate s utility function as u i = s i c i (x i + z i ) (4) where we assume that the marginal cost to raising campaign contributions is c i, whether those funds end up being reported or not. Hence, each candidate will simultaneously choose how much to spend through formal and informal channels taking into account each other s strategies. While her formal spending is capped at x, she can spend unlimited amounts informally. Let x i and z i denote the formal and informal spending of the other candidates. Her maximization problem is max s i(x i,x i,z i,z i ) c i (x i + z i ) (5) 0 x i x,z i 0 where we write the spending vectors x := (x 1,...,x I ) = (x i,x i ) and z := (z 1,...,z I ) = (z i,z i ). The solution concept we use is that of a pure-strategy Nash equilibrium: a vector of expenditure levels in which each candidate s expenditures maximizes her payoff given the expenditures of her opponents. To solve this problem, we first note that given any pair of spending vectors (x i,z i ), candidate i s marginal utility is always higher with respect to formal spending compared to informal spending. Therefore, the candidate will only spend through informal channels when she is binding at the cap. Second, given the structure of the game, candidate i s best response (x i,z i ) can be written as a function of the aggregate input of other candidates Ỹ i := k i y k. Specifically, the best response function is: (0,0) if xi 0 (xi (x i,z i ) =,0) if 0 < x i < x ( x,0) if xi x i and z i 0 ( x,z i ) otherwise [ ] [ ] where xi = a 1 ai i c i Ỹ i Ỹ i, and z i = b 1 bi i c i Ỹ i Ỹ i a i x b i. Equation 6 distinguishes between four cases. In the first, the candidate does not enter the race because the costs of doing so outweighs her benefits. In the second case, the candidate enters the race and spends exclusively through formal means some amount under the cap. In the third, she spends the exact amount of the cap through formal channels, but does not spend additional funds informally. In the fourth and final case, the candidate spends up to the cap through formal channels, and then spends on top of this through (6) 10

12 informal channels. Proposition 1. There is a unique pure strategy Nash equilibrium for the simultaneous-move subgame played by the candidates. Proof. First, rewrite the best response function ( x i (Ỹ i ),z i (Ỹ i ) ) into the input y i (Ỹ i ) chosen by each candidate as a best response of the aggregate inputs of other candidates (see Appendix A.1 for additional details). Then, transform these best response functions into share functions s i (Y ) which represent the share of total inputs that a candidate will spend as a best response when total spending by other candidates is Ỹ i Y y i. We derive this function to be { { { s i (Y ) = max min max 1 c } iy,0, a } i x,1 c } iy a i Y b i We can then sum the individual share functions into an aggregate share function: S(Y ) = I k=1 s k(y ). This function is greater than 1 for sufficiently small values of Y, equal to zero for sufficiently large values of Y, is strictly decreasing whenever positive, and is continuous. Thus, there is a unique Y such that S(Y ) = 1, which is the aggregate input in equilibrium. This value pins down the unique equilibrium spending (x i,z i ) of each candidate. (7) Comparative Statics We next consider how the spending cap x affects equilibrium outcomes. For the remainder of this section, we assume that there is at least one candidate whose formal spending is binding at the cap (otherwise, there are trivially no effects from a marginal change in the cap). For expositional purposes, we also assume that no candidate is at a knife-edge case whenever computing derivatives (i.e. we ignore the special cases x i = 0, x i = x, and z i = 0). The proofs for this section are included in Appendix A.1. Lemma 1. Total equilibrium inputs in the contest are increasing in the spending cap, i.e. Y x > 0. Proposition 2. The effects of spending limits on campaign expenditures. x i x z i x = = ( ) 1 Y a i x 1 2c iy a i if 0 < xi < x 1 otherwise [ ( ) ] 1 Y b i x 1 2c iy b i a i if z i > 0 0 otherwise While the above lemma states that total equilibrium inputs are increasing in the cap, the proposition 11

13 shows that each candidate s expenditures is not necessarily increasing in the cap. This result is an extension to previous findings in the literature studying contests and all-pay auctions in the context of political lobbying. 10 To illustrate why formal spending is not necessarily increasing in the spending cap, consider a situation where there are two high-effectiveness entrants spending at the cap and a low-effectiveness entrant spending less than the cap. Whereas the binding candidates will increase their spending with an increase in the cap, the non-binding candidate will only increase her spending if her effectiveness is sufficiently high relative to her cost of fundraising (if a i > 2c i Y ). Let us now consider a candidate who spends informally in equilibrium. A similar condition then determines whether this candidate will increase her inputs when the spending cap increases: i.e if b i > 2c i Y. Whether this translates to an increase in informal spending is less obvious, as the candidate will substitute informal spending for formal spending. If other candidates are sufficiently increasing their inputs as a reaction to the increase in the cap ( Y / x is large), it is possible for the candidate to increase both formal and informal spending. Otherwise, she will decrease informal spending because of substitution to formal spending. Proposition 3. The effects of spending limits on political entry. A candidate enters the race if and only if a i c i > Y (8) Therefore, the number of entrants in equilibrium decreases in the spending limit. We find that increasing the spending cap decreases the number of entrants. Intuitively, this is because in equilibrium, total inputs into the contest Y are increasing in the spending limit. Thus, with higher spending limits, elections are more competitive in the sense that a candidate must make more expenditures to achieve the same vote share. On the other hand, the candidate s fundraising cost is the same for any cap, and hence she is less likely to enter when the cap is high. An increase in the spending limit will also affect the composition of the pool of entrants. Equation (8) shows that the threshold to entry depends on the ratio of the candidate s popularity a i to the marginal cost c i. As the spending cap increases, the entrants with the lowest ratios will exit first. 10 Che and Gale (1998) consider a two-player all-pay auction and show that bid caps may increase total expenditures. On the other hand, considering an n-player contest, Fang (2002) finds that imposing an exogenous cap never increases total expenditures. In contrast to Fang (2002), our model also allows bidders to differ in their abilities to convert expenditures into inputs in the contest function, and hence we find that bid caps may have either effect in the n-player contest. 12

14 Suppose that a and c are uncorrelated across candidates. Then, increasing the limit will cause the entrants with the highest fundraising costs to drop out of the race. If the cost to fundraising is lower for wealthier candidates, this would result in a wealthier entrant pool. In addition, the entrant pool will be composed of more popular candidates. In this sense, only the most electable candidates will choose to run when limits are generous. If we suppose instead that a and c are correlated, then we cannot make such stark predictions. When structurally estimating the model in Section 5.5, we will flexibly allow a and c to be positively or negatively correlated in order to investigate the effects of the policy change on candidate entry. Proposition 4. The effects of spending limits on electoral outcomes. Increasing the spending limit decreases the vote share of the candidates whose equilibrium formal spending is less than the cap, and increases the vote share of the candidates whose equilibrium formal spending equals the cap. Finally, we show that an increase in the spending limit may increase or decrease an entrant s vote share. The main finding is intuitive: the candidates who spend less than the cap will face a more competitive contest under the high cap. This result has implications regarding the effect of spending limits on incumbency advantage. If incumbent characteristics are such that they are more likely to be binding spenders than challengers, then incumbency advantage will increase in the spending limit. 4 Research Design We are interested in estimating the causal effects of campaign spending limits on political entry and selection. As we discussed in Section 2, prior to the 2016 municipal elections the Brazilian government imposed a cap on the amount of money a candidate could spend in the election. The law created a discontinuous kink in the spending cap for municipalities with a candidate that spent above R$142,857 in the 2012 elections. Visually, the effects of the law on candidate spending for the 2016 elections can be clearly seen in Figure 3. For municipalities that did not have a 2012 candidate who spent above R$142,857, their candidates were capped at R$108,039. For the municipalities above this threshold, the spending cap jumps up by about 25 percent and then increases linearly as determined by the rule. It is also clear from Figure 3 that the caps were not binding for the majority of the municipalities. As a result, one should interpret our findings as intent-to-treat estimates. 13

15 To identify the effects of spending limits, we exploit the discontinuity at R$142,857 using a standard regression discontinuity design approach. Let S m,2012 denote the maximum amount spent by a candidate in municipality m during the 2012 elections. The treatment effect on outcome Y m,2016 of the spending cap is given by: Treatment Effect = lim E[Y m,2016 S m,2012 = s] lim E[Y m,2016 S m,2012 = s]. (9) s 142,857 s 142,857 The first conditional expectation measures the expected outcome at the threshold for municipalities in which candidates campaign spending is capped at R$133,700. The second conditional expectation function measures the expected outcome at the threshold for municipalities in which candidates campaign spending is capped at R$108,039. Under the assumption that these two conditional expectations are continuous in s, this difference estimates the causal effect of campaign spending limits on political outcomes, at the point of discontinuity. We estimate these conditional expectations by local linear regression using only data within a bandwidth h of the threshold. Formally, we estimate the following OLS model, for S m,2012 (142,857 h,142,857 + h), Y m,2016 = α + β 1{S m,2012 > 142,857} + δ 0 S m, δ 1 S m,2012 1{S m,2012 > 142,857} + ε m,2016 (10) where 1{S m,2012 > 142,857} is an indicator equal to 1 when S m,2012 > 142,857, and ε m,2016 represents the error term. The parameter β measures the treatment effect. For our choice of bandwidth h, we rely on the approach developed by Calonico et al. (2014). This optimal bandwidth choice is a function of the data and is thus different for each outcome, Y m,2016. We also explore the robustness of our results to alternative bandwidth sizes. Before presenting our results, it is important to test the validity of our research design. In Panel (a) of Figure 4 we plot the density of our running variable, S m,2012. Unsurprisingly, we do not find any evidence of manipulation or endogenous sorting around the discontinuity threshold. This is completely expected: campaign expenditures are made public immediately following each election, and no one could have anticipated the recent law change back in As a point of comparison, Panel (b) of Figure 4 plots the distribution of campaign spending for the 2016 election. In contrast to the previous plot, Panel (b) does exhibit substantial bunching at the spending cap of R$108,039. Another general concern associated with regression discontinuity designs is the possibility that other determinants of our outcomes of interest are also varying discontinuously at the cutoff point. Although we cannot directly test this assumption for unobserved characteristics, we can examine 14

16 whether any observable characteristics of the municipality also exhibit discontinuous jumps at the cutoff point. In Figure 5, we present a series of plots, exploring various municipal characteristics that are correlated with our political outcomes of interest, such as GDP per capita, illiteracy, and the share of the urban population. 11 In each graph, we plot a bin scatter of the municipal characteristic against the maximum amount a candidate spent in the municipality during the 2012 elections (i.e. our running variable). In addition to these binned averages, we also fit a second-order polynomial on each side of the point of discontinuity and 95% confidence intervals for each bin. We do not find any evidence of other characteristics jumping at the cutoff point. All the differences are close to zero in magnitude and statistically insignificant. Importantly, these comparisons also include our main political outcomes of interest but measured for the 2012 elections (i.e. the pre-treatment period ). 5 Results 5.1 Effects of Spending Caps on Campaign Expenditures and Contributions In this section, we estimate the causal effects of the spending caps on candidates campaign spending and contributions. We begin with the graphical evidence. In Panel A of Figure 6, we plot binned averages of the amount candidates spent in the 2016 elections against our running variable (the maximum amount spent by a candidate in the 2012 elections centered at R$142,857). We also fit a second-order polynomial, separately estimated on each side of the discontinuity. The discontinuity at zero provides an estimate of the gap in candidates campaign spending imposed by the law. The estimated discontinuity implies that a 25 percent increase in the spending cap increased maximum campaign spending by approximately 12 percent during the 2016 elections for municipalities near the discontinuity. In Panel B, we reproduce the graph presented in Panel A, but for the mean amount spent by a candidate. We see a similar increase of approximately 10 percent, which further suggests that the caps did bind for many candidates. We refine the graphical analysis in Table 3. Each row corresponds to a different dependent variable, and each numbered column presents the estimated impact for a different regression specification. In column 1, we present our baseline estimates of Equation 10, using the bandwidth proposed by 11 These plots represent only a subset of the characteristics for which we tested. Table 2 presents the entire set. Out of the 18 municipality characteristics tested, only one displayed a discontinuous jump at the cutoff point (population). For this reason, we control for the municipal characteristics measured in the 2010 Census in our analysis. 15

17 Calonico et al. (2014). In columns 2 and 3, we explore the robustness of our estimates to different bandwidth choices. In column 4, we further test the sensitivity of our results by fitting a local quadratic polynomial on each side of the discontinuity instead of a local linear polynomial. Our results are robust to these various modeling choices. In our baseline specification, the highestspending candidate just to the left of the discontinuity spent on average R$84,823 to become mayor, compared to R$95,036 for candidates in municipalities just to the right of the discontinuity. This represents a 12.0 percent increase in spending. The point estimates in columns 2-4 are similar: they indicate increases in maximum spending ranging from 11.6 to 13.9 percent. The estimates on average spending, although a bit noisier, are also consistent across specifications. They imply that the higher spending cap led to increases in mean spending ranging from 8.5 to 11.7 percent. The theory does not provide clear predictions on the effects of a spending cap on the minimum or total amount spent in an election. In some cases, a higher spending cap will induce the minimumspending candidate to reduce spending further, or even exit the race. Thus increasing the spending cap does not necessarily lead to an increase in the minimum or total spending within a race. Consistent with the theory, we do not find significant effects on either of these two outcomes. The minimum amount spent by a candidate is similar on both sides of the point of discontinuity: we estimate a statistically insignificant increase of R$989. Similarly, we also find a statistically insignificant increase in total spending of about 4 percent at the cutoff point. In Table 4, we consider the effects of spending caps on the amount and composition of the candidates campaign contributions. On average, candidates spend 99% of their campaign contributions. 12 Reflecting our findings on spending, we find that the average amount of campaign contributions raised by candidates are R$6,179 higher for municipalities with the higher limit. Approximately 75% of this increase comes in the form of candidates financing their own campaigns, which likely stems from the law s ban on corporate donations. In 2012, candidates received on average 16 percent of their contributions from corporations, and self-financed 25 percent of their campaigns. In contrast, 2016 candidates self-financed 40 percent of their campaign expenditures. We can interpret these results in two ways. On the one hand, in the face of the corporate ban, the higher caps induced the existing candidates to contribute more to their own campaigns. On the other hand, higher caps may have attracted a wealthier pool of candidates with greater financial means to run for office. We examine this possibility in the next section. 12 In Brazil, candidates are not allowed to accumulate war chests. 16

18 5.2 Do Spending Caps Affect Candidate Entry? According to our model, the number of candidates who enter the race should decrease as spending limits increase. Additionally, higher spending limits may also attract individuals who have a higher ex-ante probability of winning. We test these predictions in Table 5. As before, the rows indicate different dependent variables, and the numbered columns present the estimated effects of the caps for different modeling choices. Spending caps affect the entry decisions of potential candidates. Compared to the municipalities just to the right of the threshold (i.e. the less constrained municipalities), the cap led to a 0.26 increase in the number of candidates for municipalities capped at R$108,039. On average, 2.9 candidates run for mayor, so this effect represents a 9 percent increase in the size of the candidate pool. This result is presented visually in Figure 7. In contrast to the plot presented in Panel C of Figure 5, which displayed the effects on the number of candidates who participated in the 2012 elections, we see a significant jump in the number candidates at the point of discontinuity. To test whether this increase in candidate entry actually increased political competition, we study the effect of the cap on the effective number of candidates. This measure is computed by taking the inverse of the sum of squared vote shares of each running candidate within an electoral race. If all candidates have the same vote share, then this measure is equal to the actual number of candidates. At the other extreme, if one candidate wins every vote, then the effective number of candidates is one. If a change in the spending cap only leads to the entry or exit of candidates winning few or no votes, then we would not find an effect on the effective number of candidates. On the contrary, we find that the more restrictive spending cap increases the effective number of candidates by 0.143, suggesting that the restrictive cap did increase the competitiveness of mayoral races. In Table 5, we also test whether higher caps affected the types of parties that entered the contest based on their size and ideology. To measure ideology, we rely on a measure of party position along a left-to-right scale as created by Power and Zucco (2012). The index, which ranks parties from 1 (= left ) to 10 (= right ), is constructed from a survey of federal legislators elected in We find no evidence that the caps impacted the average ideological score of the candidate pool, nor the tails of the distribution. The increase in political competition was also not entirely the result of smaller parties entering into the race: higher caps reduced the number of smaller parties by 6 percent, although the effects are imprecisely measured We define the small parties to be all political parties except for the six most successful in the 2012 municipal elections: the PMDB, PP, PSB, PDB, PSDB and PT. Together, these six parties won the majority of mayoral elections in In total, there are thus 30 small parties in the 2016 elections. Our results are robust to the choice of party 17

19 To examine whether higher spending limits also induce greater participation from individuals with a higher ex-ante propensity to get elected, we first estimate the probability of winning the 2016 election based on the follow set of observable characteristics: gender, age, race, education level, political experience, party affiliation and self-reported assets. We estimate this propensity score for the sample of candidates that are outside the bandwidth of the RD regressions. The results, which are reported in Table A.1, suggest that candidates who are male, wealthier, incumbents, or have more political experience are more likely to win. Based on these estimates, we then impute a candidate s ex-ante probability of winning the election. We find that individuals with higher expected winning probabilities are more likely to participate in municipalities with a higher spending limit. For a 25 percent increase in the spending limits, high-propensity types are 2.0 percentage points more likely to enter, which represents a 6 percent increase. To see where these effects are coming from, in the remaining rows, we estimate the effects on individual attributes of the candidate pool. Although the estimates tend to be fairly noisy across the various attributes, higher limits do appear to affect an important factor: they tend to attract wealthier candidates. In our baseline specification, the average level of assets among candidates is 40 percent higher in municipalities with a higher spending cap. This result is perhaps unsurprising given our finding that the majority of the extra spending under the high-cap is self-funded. 5.3 Spending Caps and Political Selection While restricting campaign spending does increase political competition, it appears to do so at the cost of attracting individuals with a lower ex-ante propensity to be elected. Whether spending caps affect political selection is therefore an empirical question. The graphical evidence presented in Figure 8 suggest that it does. Here, we plot binned averages of re-election rates against the maximum amount spent in the municipality by a candidate for the 2012 elections. In computing this graph, we restrict the sample to the 2,721 incumbents who were eligible for re-election. We see a positive jump in the reelection rate at the point of discontinuity. In Table 6 we refine the analysis further, by considering a range of alternative specifications and by conditioning on whether or not the mayor ran for re-election. In Panel A we consider all incumbents who are not term-limited, whereas in Panel B we consider only those who run for re-election. When classification. 18